Bullet Validation

Status: Current Last updated: 2026-10-20 (commit f7e32e22)

Media bullets link transcript positions to unsigned millisecond offsets. The visible •start_end• notation in this chapter stands for actual U+0015 delimiters. Parsing, interval validation, document ordering and acoustic accuracy are separate questions. This chapter describes Chatter’s current checks, not a claim that its timing matches the recording.

Current policy by timing surface

Bullets encode unsigned millisecond offsets between U+0015 delimiters; • is the visible notation used in this chapter. Parsing a bullet is not proof that its interval is valid, nor that it matches the recording.

Equal starts pass the start-order check; touching intervals and gaps are not errors. Exactly 500 ms of same-speaker utterance overlap passes E704. The word-level policy does not inherit that utterance-level tolerance: it has no adjacent-word overlap test at all. For example, %wor intervals 100_300 and 200_400 are both positive and allowed, as are 300_400 followed by 100_200. These are structural permissions, not acoustic endorsements.

Important scope distinction: the global E362 check in ChatFile::run_validation_checks still runs alongside the per-speaker E701 check. Thus a cross-speaker backward start can pass E701 but fail complete validation with E362. Statements below about E701’s per-speaker scope must not be read as permission for global backward starts. Start ordering and end overlap are different relationships; permitting conversational overlap does not disable the start-order requirement.

%wor has no trailing tier-level bullet. Its timing belongs to individual words; its main tier owns the utterance span. See %wor Timing Semantics for membership, lexical corroboration, complete timing admission, and source-bound replacement. Zero-duration word bullets are now invalid even in older data; this rule does not rewrite it. The special 0_0 spelling can already fail parsing as E360, before interval validation; it is not an accepted exception to positive duration.

Media linkage and evidence

An utterance-final bullet, an internal main-tier bullet (including nested content), or an actual %wor word bullet supplies timing evidence. A %wor tier containing only untimed words does not. Timing presence does not depend on lexical alignment or matching word counts.

• E752: timing exists but no @Media header declares the indexed media.
• E552: @Media says unlinked but timing exists, including %wor-only timing.
• E544: an expected recording is declared without a status qualifier, but no timing exists. A declaration of missing media is not a linkage promise.

Replacement admission may carry an E544 regeneration obligation while a designated timing tier is removed. It must be discharged before complete output admission; removing bad timing is not proof of successful regeneration. None of these checks opens audio to certify recording length or acoustic placement. Validation reports faults; it does not manufacture durations or normalize overlapping words.

Which bullet does each check use?

The global E362 sequence consists only of explicit utterance-final bullets, in document order. Untimed utterances are skipped; the comparison is between successive entries in that selected sequence, even if another utterance lies between them.

The temporal E701/E704 sequence instead uses Utterance::media_bullet(). For each utterance it prefers the explicit final bullet; if none exists, it uses the first top-level InternalBullet. This fallback does not recursively collect nested bullets and does not use %wor. Each speaker’s successive entries are compared independently, including across intervening speakers.

Media-linkage checks use a broader timing-evidence view, including nested main-tier and word-tier timing. Do not substitute any one of these collections for another merely because each contains bullets.

The following flow separates the document-order and speaker-order checks. The owners are model/file/chat_file/validate.rs, model/file/utterance/accessors.rs, validation/bullet.rs and validation/temporal.rs in talkbank-model.

flowchart TD
    F["ChatFile::run_validation_checks"] -->|"explicit final bullets"| G["check_bullet_monotonicity"]
    G -->|"a start decreases in document order"| E362["E362"]
    F -->|"validate_temporal_constraints"| C["collect_bullets: Utterance::media_bullet"]
    C -->|"per-speaker prior start"| S["validate_global_timeline"]
    C -->|"per-speaker prior end"| O["validate_speaker_timelines"]
    S -->|"current start is earlier"| E701["E701"]
    O -->|"prior end minus current start exceeds 500 ms"| E704["E704"]

Despite its historical name, validate_global_timeline implements the per-speaker E701 check, not global E362. Each temporal pass updates its speaker state after every selected bullet, including when it reports an error. Neither temporal pass short-circuits the other. E362, E701 and E704 can therefore coexist on one input.

Every selected interval participates, including utterances containing only www, xxx or yyy. @Options: CA does not bypass these checks. The separate lexical-content predicate is not a timing exemption.

E704 has two independent producers

Timed self-overlap compares the current selected bullet’s start against the same speaker’s previous selected bullet’s end. More than 500 ms reports E704; exactly 500 ms does not. This is a comparison with the most recent selected interval, not every earlier interval or a maximum-end envelope.

A separate cross-utterance marker check also reports E704: adjacent utterances have the same speaker, the first has a paired top overlap region and the second has a paired bottom region. An ordinary same-speaker continuation without that bottom region is not enough. This path needs no bullet and has no millisecond tolerance. See the E704 spec’s paired marker controls.

The diagnostic code alone therefore does not identify which relationship failed. Read its message and source location before proposing a correction.

CHECK comparison: use the retained contract

The generated CHECK assessment comes from the checked parity manifest and its retained fixtures. Its finite claims are the comparison authority:

E701 is an additional per-speaker start-order diagnostic; a numeric mapping to CHECK 83 does not make their scopes identical. The manifest also records fixture-specific CHECK behavior, including the three-tier witness needed for 83. Do not infer general diagnostic counts or intent from those observations.

E729 (cross-speaker overlap), E730 and E732 remain reserved and unimplemented, not implemented checks waiting for an option. E731 does not add a second zero-tolerance self-overlap rule; the E704 timing check owns that relationship. The E729 and E731 specs retain default-mode overlap and exact-500-ms controls.

This chapter does not claim a current CLAN source audit, identical recovery, identical diagnostic counts, or an implementation fault in another validator.

Reviewing a timing finding

1. Identify the surface: final main-tier interval, fallback internal interval, word interval, or overlap-marker relationship.
2. Read the diagnostic message. E362 covers both individual intervals and global start order; E704 covers timing and marker relationships.
3. Compare against the appropriate predecessor and source span, not merely the preceding printed line.
4. Review the recording and transcription intent before editing. Do not sort turns, clamp times or remove bullets simply to silence a diagnostic.
5. Validate the reviewed result. Passing checks establish the stated structural rules, not acoustic correctness, complete word timing or media provenance.

For consumers needing complete positive word intervals and a timing hull, use the checked transitions in %wor Timing Semantics. Ordinary validation permits partially timed word tiers and does not grant that stronger capability.

This page last changed: 2026-10-20 (commit f7e32e22). The whole book last changed: 2026-10-27 (commit 5bfee2f9).